Automatic recognition for tunnel lining voids in GPR images based on SC-PGGAN and bidirectional collaborative data augmentation

The quality of tunnel lining is a key factor affecting the quality of tunnel construction. Currently, lining defect identification relies on manual interpretation of ground-penetrating radar (GPR) data, which is inefficient and its accuracy depends on the experience of on-site personnel. Additionally, developing high-accuracy automatic detection models requires a large number of data samples, but it is difficult to obtain sufficient samples in practical research. Existing data augmentation methods suffer from issues such as loss of details in GPR images and low image quality. To address this issue and improve tunnel lining void identification performance, this paper proposes a Spectrum-Constrained PGGAN (SC-PGGAN) network model, which can generate a large number of high-quality GPR images. Experiments show that the images generated by the SC-PGGAN model outperform the GAN and PGGAN models in terms of FID, PSNR, and SSIM metrics. Furthermore, a method for automatic identification of lining voids based on bidirectional collaborative data augmentation is proposed to improve void detection accuracy. Experiments show that training the YOLOv13 model with this method achieves an mAP@50 of 91.70%, outperforming Faster R-CNN (76.04%), RT-DETR (82.54%), and YOLOv13 (84.50%). This paper provides an effective method for high-precision identification of tunnel lining voids.